Stimulus Selection

LAION-fMRI stimuli were selected to cover a broad range of natural image content while preserving enough repetitions for reliable single-trial fMRI estimates. Most images come from LAION-natural, a curated 120-million-image subset of LAION-2B filtered to natural photographs (Roth & Hebart, 2025). The set is supplemented with images from NSD and THINGS / THINGSplus for cross-study comparison, plus a shared out-of-distribution (OOD) set for generalization tests.

The launch-release stimulus set contains 25,052 distinct images across five participants. Each participant saw 6,204 unique images: 1,492 shared images common to all participants and 4,712 subject-unique images. Supplemental sessions expand the shared set to about 2,200 images and will be released later.

Selection Goals

The stimulus set was designed around three goals:

• Broad coverage of the natural-image distribution, rather than a small hand-picked taxonomy of categories.

• Cross-study comparability through overlap with established neuroimaging image sets, especially NSD and THINGS / THINGSplus.

• Evaluation of generalization through predefined train/test splits and a shared OOD image set.

The resulting dataset has no designed semantic category taxonomy. Coverage is driven by feature-space diversity, and category-like metadata should be treated as derived annotations rather than as the sampling frame.

Source Pools

Source	Role	Notes
LAION-natural	Main source pool	Natural photographs selected to cover CLIP feature space broadly.
NSD	Shared comparison set	Supports comparison with the Natural Scenes Dataset.
THINGS / THINGSplus	Object-image comparison set	Supports comparison with THINGS and THINGS-data EEG + fMRI.
OOD	Held-out stress test	Visual illusions, Gabor patches, shapes, cropped textures, unusual configurations, self-made images, and high-saturation patterns.

Selection Procedure

The LAION-derived images were selected in two stages.

First, a large LAION-natural sample was embedded with CLIP, reduced with PCA, and clustered. Candidate images were then chosen to maximize the effective dimensionality of the selected set, encouraging broad coverage of independent axes in visual-semantic feature space.

Second, visually similar alternatives were retrieved around each selected prototype using approximate nearest-neighbour search. These candidate pools made it possible to replace rejected images while preserving the same broad feature-space coverage.

Images assigned to the shared set were the feature-space slots chosen for repeated measurement in every participant; the remaining accepted prototype-centred pools supplied the unique set by contributing one quality-approved, non-overlapping candidate image per participant.

NSD and THINGSplus images were integrated by assigning them to nearby LAION feature-space slots. When an NSD or THINGSplus image filled a slot, it replaced the corresponding LAION shared image so that the shared set size stayed fixed while gaining cross-study overlap.

Manual Screening

Candidate images were reviewed in a custom web interface. Reviewers accepted clear natural photographs and rejected images with applied filters, illustrations, graphic-design content, severe quality problems, or not-safe-for-work content. Candidate pools likely to contain NSFW material were flagged before manual review using an automated caption-based screen.

If a prototype pool did not have enough accepted images for all participants, additional neighbours or newly selected LAION-natural candidates were reviewed and added.

Final Counts

Shared set (1,492 images viewed by every participant)

Source	Count	Repetitions
LAION-natural	641	12
THINGSplus	240	12
NSD	240	4
OOD	371	4

Each participant additionally viewed 4,712 subject-unique images: 4,246 from LAION-natural, 144 from THINGS, and 322 from THINGSplus. Across all five participants, this yields 25,052 distinct launch-release images.

See Stimulus Set for file formats, metadata fields, and loading examples. See Train / Test Splits for the predefined evaluation splits.